As Internet of Things (IoT) devices become increasingly interconnected, they are exposed to cybersecurity risks. A particularly critical threat is posed by poisoning attacks, where adversaries deliberately inject harmful gradients into the training process, thereby compromising the reliability and accuracy of the learned model. Existing detection methods attempt to mitigate this threat, but they often struggle to process the vast and heterogeneous data generated by IoT systems. While deep learning solutions have shown promise, they typically rely on centralized datasets, limiting scalability and effectiveness. Federated Learning (FL) has emerged as a compelling alternative, enabling decentralized model training without raw data sharing. However, FL remains highly vulnerable to distributed data poisoning, which can severely compromise the global model. In this work, we propose the Adaptive Federated Learning Approach for Detecting Poisoned Data(AFA-DPD). Our method introduces a proactive filtering layer that leverages the Isolation Forest algorithm to identify and exclude suspicious clients before training phase. Experimental results demonstrate that AFA-DPD enhances the robustness of recent state-of-the-art FL systems when combined.

Graph Neural Networks (GNNs) are a revolutionary game-changing approach toward psychiatric diagnosis because of their incomparable capability for modeling complex relations in neuroimaging data. Herein, we introduce an explainable high-powered GNN-based model designed to address the challenge of distinguishing patients with Major Depressive Disorder (MDD) from healthy t method’s foundation is on the following new suggestions: feature extraction, advanced hyperparameter adjustment, and powerful explainable GNN (X-GNN). Our model, tested on the Rest-Meta-MDD dataset, demonstrated exceptional performance while achieving state-of-the-art performance.

Predicting energy consumption plays a crucial role in promoting energy conservation and reducing power generation costs. Recent research indicates a growing interest in utilizing machine learning algorithms for predicting energy consumption in households. This research utilizes a massive dataset of several millions, containing electricity consumption records of residential households in Uruguay (mostly in Montevideo). In this study we examine the utilization of multiple machine learning techniques such as linear regression, random forest, extra-trees regressor, lasso regression, xg boost, and ridge regression to predict household electricity demand. These models were trained and evaluated using historical electricity usage data. In order to evaluate these models, the coefficient of determination (R squared) metric is employed. Tree-based algorithms, including random forest, extra trees regressor, and xgboost, achieve the best results. Among them, random forest demonstrates the highest performance.

Detection of mental health conditions like anxiety, depression, and post-traumatic stress disorder (PTSD) at their early stages is crucial for successful treatment interventions. This paper introduces an innovative framework that combines multiple data modalities for identifying psychological disorders. Our approach synthesizes three distinct data sources: audio recordings of speech, text transcriptions of conversations, and clinical measurements including PHQ-8 questionnaire results and patient demographics. At the core of our methodology is a sophisticated attention-driven fusion system that intelligently calibrates the influence of each data stream according to individual patient contexts, generating a comprehensive representation of their mental state. To support clinical understanding, we implement explainable artificial intelligence methodologies (SHAP) that highlight key contributing factors in the classification process and offer healthcare providers meaningful insights into the model's decision-making. The system demonstrated exceptional performance with 95.83% accuracy in differentiating between anxiety, depression, and PTSD cases, surpassing previous approaches that relied on single data types.

Stroke continues to be a major contributor to mortality and disability, highlighting the need for precise early detection systems. This research introduces an innovative combined approach utilizing Fuzzy logic and Long Short-Term Memory (LSTM) networks, enhanced with Explainable AI (XAI) methodologies, to evaluate stroke risk through electroencephalography (EEG) data analysis. Our methodology leverages LSTM networks' temporal pattern recognition capabilities alongside a fuzzy inference framework that converts medical expertise into comprehensible linguistic guidelines. We incorporate XAI tools— SHAP—to overcome deep learning's opacity by providing comprehensive explanations at both global and individual prediction levels. The system architecture processes EEG characteristics and patient clinical information through separate LSTM channels, then integrates these outputs with a fuzzy evaluation system to produce understandable risk assessments. Testing on clinical EEG information yielded remarkable predictive performance: 98.79% accuracy with the Ensemble LSTM and perfect accuracy with the hybrid Fuzzy-ELSTM approach.

This paper proposes a single-objective filter pruning method based on constrained binary Particle Swarm Optimization, integrating L1-based search space reduction and adaptive weighting. The proposed method, AWL-BPSO, ensures efficient convergence while maintaining a diverse set of pruned models, providing multiple trade-offs between accuracy and computational efficiency. Unlike existing methods that rely on fixed pruning rates, our method dynamically explores optimal filter configurations, making it more flexible for real-world deployment on edge devices. Experimental results on popular deep convolutional neural networks demonstrate that our method achieves significant model compression while preserving accuracy, outperforming conventional multi-objective methods in both efficiency and adaptability. This work contributes to advancing scalable, efficient, and hardware-aware neural network pruning for deep learning applications.

This research proposes a Federated Multi-Modal Deep Learning Framework (FedMM-AD) with explainability features and privacy-preserving techniques for early-stage AD prediction. Our federated model was constructed using MRI, PET, and CT scan data, protecting patient privacy and security while enabling cross-institution training of the model without sharing any patient information. The FedMM-AD model’s accuracy, AUC, sensitivity, and specificity were 98.59%, 97.3%, and 98.2%, respectively, according to the experiment conducted on the ADNI and OASIS datasets. In order to increase the model’s transparency and help doctors identify key brain regions for AD, we integrated explainable AI techniques employing cross-attention. The findings show that when additional stakeholders, like clinicians, are properly involved, the approach can be beneficial and have an impact.

Electrocardiogram (ECG) and electrocardiography are essential tools for monitoring heart activity and diagnosing cardiovascular diseases. However, the accuracy of ECG signals can be compromised by various types of noise, including environmental factors, patient movement, and electrical interference. Among the most significant noise sources affecting ECG signals are Baseline Wander (BW), Electrode Movements (EM), Motion Artifacts (MA), Additive White Gaussian Noise (AWGN), and Powerline Interference (PLI). This paper provides a detailed review of the challenges posed by these noise sources and explores advanced techniques for minimizing their impact on ECG signals. By employing signal processing methods such as filtering, adaptive algorithms, and wavelet transforms, autoencoders, and generative methods, these techniques aim to improve the reliability and accuracy of ECG data analysis. Enhanced noise reduction enables healthcare professionals to make more precise diagnoses and treatment decisions, ultimately contributing to better patient outcomes in cardiovascular care.

Facial detection plays a pivotal role in computer vision applications, spanning facial recognition, surveillance, and augmented reality. This study delves into the efficacy of the Hybrid method, combining the Histogram of Oriented Gradients (HOG) algorithm with a Support Vector Machine (SVM) classifier for robust facial detection. HOG emphasizes extracting pertinent features, focusing on the shape and texture of objects within images. These features serve as inputs for an SVM classifier, training it to distinguish faces from other objects. The research aims to create a precise facial detection system adaptable to diverse conditions, encompassing variations in orientation, lighting, and shadows. The emphasis lies on optimizing the HOG-SVM approach to achieve exceptional performance without compromising processing efficiency. Experimental evaluations, conducted on the LFW and Wider datasets, demonstrate promising results, showcasing the superior accuracy of our HOG-SVM approach compared to recent methodologies.

Brain tumors, whether benign or malignant, present significant challenges in medical diagnosis and treatment. Timely and precise detection is critical for effective intervention and patient outcomes. This study introduces a pioneering method for brain tumor detection, employing a fusion of Convolutional Neural Networks (CNN) and Vision Transformer (ViT) architectures. By integrating these models, we exploit their complementary features in image analysis, particularly in medical imaging contexts. Our research assesses the performance of this integrated CNN-ViT framework across various brain tumor imaging modalities and clinical scenarios using extensive experimentation on benchmark datasets. Results validate the robustness and accuracy of our approach, achieving a remarkable precision, recall rates, and overall accuracy of 98%.

The fundamental concept behind image steganography is to conceal one image within another. To advance steganography, it may be wise to conceal multiple images within other multiple images. The major goal of our suggested strategy is to conceal a collection of related images while taking into account size equality. With the aid of a 3D-DeepCNN grounded autoencoder, we introduce a novel multi-image steganography approach in this study. Within four cover images, we attempt to encode and decode four secret images. The quantitative findings show that the suggested model shares the embedded hidden image information over every component of the cover image, without compromising image quality. The results of our model using the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are (28.56, 0.928) for global secret input and (32.516, 0.973) for global cover input. The qualitative results were evaluated and give an effective performance in comparison to the current models.

Brain tumors are a pathological condition characterized by aberrant growth within the cerebral structure. The segmentation of these tumors becomes pronounced to discern their boundaries accurately amidst healthy brain tissues, owing to the variability in tumor shapes and the complexities of determining their location, size, and texture. Manual tumor segmentation, a time-consuming task, is highly susceptible to human error. In this paper, we propose a deep inception residual network for brain tumor segmentation using the UNet architecture with a pre-trained Inception ResNet V2 encoder. The Inception-Resnet block, which fuses Inception with the residual neural network, is included. With this architecture, segmentation is significantly more robust. The study is being carried out on the RSNA-MICCAI Brain Tumor Radiogenomic Classification data set, i.e., on the benchmark dataset BraTS 2020. Our network achieves dice scores of 85.7%, 91.2%, and 83.2% for enhancing tumor, whole tumor, and tumor core, respectively. The experimental findings demonstrate the effectiveness of our proposed approach compared to other methods.

Electrocardiogram (ECG) signal analysis is a crucial tool for enhancing the efficacy of clinical diagnosis, particularly in detecting arrhythmias. However, its performance tends to degrade under the inter-patient paradigm, especially for minority sample categories. To address this issue and enhance the detection performance of these less represented classes within the inter-patient test protocol, this paper proposes a novel hybrid framework. This framework integrates specialized blocks from convolutional networks, namely DeepArr CNN, with sequence-to-sequence BiLSTM models. The proposed approach involves extracting local ECG features from a sequence of heartbeats utilizing DeepArr CNN. These extracted feature maps are then fed into the RNN encoder-decoder to facilitate fusion with the feature maps of neighboring heartbeats. Our proposed method, which adhered to the AMII standard of the MIT-BIH arrhythmia database and operated within the inter-patient paradigm, achieved impressive accuracy rates. Specifically, it yielded accuracy rates of 99.92% and 99.81% for three and five class scenarios, respectively. The performance measures across different classes are as follows: for the N class, sensitivity (SEN) is 99.81%, positive predictive value (PPV) is 99.73%, and specificity (SPEC) is 97.81%; for the S class, SEN is 92.97%, PPV is 96.01%, and SPEC is 99.85%; for the V class, SEN is 99.97%, PPV is 99.26%, and SPEC is 99.95%; and for the F class, SEN is 97.42%, PPV is 95.70%, and SPEC is 99.97%. Even when dealing with an unbalanced dataset, our proposed solution consistently yields remarkable outcomes. Specifically, it achieves an accuracy rate of 99.24% in the three-class scenario and 98.75% in the five-class scenario. These experimental results underscore the effectiveness of our S2SDeepArr model for ECG Classification under the inter-patient paradigm, as demonstrated across various experimental cases.

Word Embeddings (WE) play a crucial role in capturing the meanings of words. They provide continuous vector representations that encode semantic and syntactic information. To accurately convert words into meaningful vectors, in this paper, we propose a novel approach called Soft EMBedding method (SoftEMB). SoftEMB combines the strengths of the Glove and Word2Vec methods through a Soft-Voting algorithm. The SoftEMD approach aims to improve the performance of word embedding, particularly in the context of Sentiment Analysis (SA) hybrid models. To evaluate the SoftEMD performance, we test it on various SA models based on CNN-LSTM, CNN-GRU, CNN-biLSTM, and CNN-bi-GRU. Our results demonstrate a substantial enhancement in accuracy when evaluating movie reviews, with scores of 88.29%, 88.33%, 88.27%, and 88.27%. Similarly, for Sentiment140 dataset, our proposal shows improved results, achieving accuracy rates of 83.27%, 82.78%, 82.76%, and 82.51%. These results highlight the significant progress made in accurately analyzing both movie reviews and the Sentiment140 dataset.

Malaria disease should be considered and handled as a potential restorative catastrophe. One of the most challenging tasks in the field of microscopy image processing is due to differences in test design and vulnerability of cell classifications. In this article, we focused on applying deep learning to classify patients by identifying images of infected and uninfected cells. We performed multiple forms, counting a classification approach using the HSV color space. HSV is used since of its superior ability to speak to image brightness, at long last, for classification, a convolutional neural network (CNN) architecture is created. Clusters of focuses were used to deliver the classification. The highlights gotten to be forbidden, and a few more clamor sorts are included to the information. The suggested method has a precision of 99.79%, a recall value of 99.55%, and provides 99.96% accuracy.

This paper presents a novel method to enhance the predictive performance of the Echo State Network (ESN) model by adopting reservoir topology learning. ESNs are a type of Recurrent Neural Network (RNN) that have demonstrated considerable potential in various applications, but they can be challenging to train and optimize due to their random initialization. To improve the learning capabilities of ESNs and enhance their effectiveness in a broad range of predictive tasks, we utilize a structure learning algorithm. The proposed approach modifies the ESN reservoir's connectivity by applying techniques such as reversing, deleting, and adding new connections. We evaluate our proposal performance using both synthetic and real datasets, and our results indicate that it can substantially improve predictive accuracy compared to traditional ESNs.

Echo State Network (ESN) is a typical version of the Recurrent Neural Network model (RNN) which is characterized by sparse reservoir and simple linear output. It has been utilized in several applications, especially for time series forecasting. Nonetheless, the ESN has some drawbacks that are mainly related to the reservoir properties and initialization (weights and connection). Thus, creating an efficient ESN model represents a challenging task. Relying on the initial structure of ESN, we propose an improved version called Dual Input-ESN (DI-ESN). This work aims to optimize the prediction error. Experimental results demonstrate that the DI-ESN model outperforms existing improved ESN models in terms of prediction accuracy.

This paper presents a novel learning algorithm formulated as a classification problem with smartphone-inertial measurement as predictor. For this reason, we will focus on the Symbolic Aggregate approXimation (SAX) as a low-cost classification algorithm resulting from its ability of the data dimensionality reduction through symbolic discretization. Since the first publication of the SAX, a lot of extension with novel SAX-distance measure are published. Each of them attempts to integrate additional statistical features in order to improve original SAX average-based feature. However, none of them can fit the overall shape-characteristics of the data and give the superiority to an individual SAX-based classifier. In order to combine the prediction of each single SAX-based classifier, we propose a collection of several SAX-feature to compose a multi-SAX feature ensemble classification. The proposed algorithm is validated on experimental data addressing (i) the character recognition when using a smartphone as a pen and (ii) the road profile estimation using a smartphone mounted in the interior of a vehicle while driving.

Automated sensing of the student's engagement in an e-learning system from emotional expressions remains a challenging problem due to varying conditions during the lecture. Such recognition and detection systems improve the teaching experience and efficiency by providing valuable feedback. Emotional expressions are expressed through non-verbal and verbal human emotional/behavior. More investigations are needed in this domain to carry out the learning process. Deep multi-task learning has been successfully employed in many real-world large-scale applications such as recognition systems. In this paper, we propose a novel education level state system to determine the student engagement level in an e-learning environment. The proposed approach is based on a hybrid deep multi-task learning technique. Soft and hard parameters are fused to achieve the best prediction. The performance of this system is evaluated on three facial expression benchmark datasets acquired in non-controlled environments. We validate the proposal using multi-input and mixed data to meet the relevant challenges.

Classification methods are carried out in several steps. The most important step is the development of classification rules based on a priori available knowledge; this is the learning phase. This phase uses either deductive or inductive learning. Inductive learning algorithms derive a set of classification rules (or standards) from a set of already classified examples. The goal of these algorithms is to produce classification rules to predict the assignment class of a new case. Among the available knowledge we can mention the choice of the initial cluster centers which is a very important factor in the final definition of clusters. For this purpose, we proposed an evolutionary algorithm EK-means based on the combination of k-means and fuzzy c-means by touching the initialization phase of the centroids. the performance of EK-means is compared with k-means according to the metrics of interclass and intraclass distances. To compare the efficiency of our optimization solution with traditional K-means, we rely on a UCI machine learning repository. The comparative study indicates a remarkable efficiency of our proposal, regardless of the type of data.

Echo State Network (ESN), as a paradigm of Reservoir Computing (RC), refers to a well-known Recurrent Neural Network (RNN). Its randomly generated reservoir represents the main reason for its ability of rapid learning. Nevertheless, designing a reservoir for a specific role constitutes a difficult task. To resolve the challenge of the reservoir structure design, in this paper, a new combination of two optimization methods, Particle Swarm Optimization (PSO) and Stochastic Gradient Descent (SGD), have been proposed to reach a higher performance. The resulted model was tested using Mackey Glass and NARMA 10 benchmarks. The experimentations proved that the suggested PSO-SGD-ESN model performs well in time series prediction tasks and outperforms the original one.

Recognizing human expression is one of the most popular problems in the Human-Computer Interaction field. Facial Expression Recognition present a great challenge in a wide variety of areas due to varying conditions of the image, which influences expression recognition and makes this task a complex problem. The main difficulties depend on the irregular nature of the human face and the different conditions such as orientation, light and shadows. Lately, Deep learning obtained more attention as an intelligent technology to achieve robustness and offer best performance of expression recognition. Further investigations are still needed in this field in order to make the recognition process very efficient. For that, we present in this paper a new Convolutional Neural Networks model enhancing with pre-processing stages to recognize seven classes (six basic expressions and one neutral). Our approach contains two phases: normalization, and expression recognition. The result can achieve high accuracy compared to recent works with the popular facial expression databases such as CK+, JAFFE, and FER-2013.

Abstract The evaluation of visual analytics (VA) is a challenging field enabling analysts to get insight into diverse data types and formats. It aims at understanding events described by data and supporting the knowledge discovery process by integrating different data analysis methods. Recently, the evolution of intelligent decision support systems has enabled the inductive and predictive approaches of data analysis to make important decisions faster with a higher level of confidence and lower uncertainty. This paper introduces a new and intelligent evaluation method of VA that understands the users’ work as well as the features of their environments including vagueness, uncertainty and ambiguity due to workload. To this end, we apply an adaptive neuro-fuzzy inference system (ANFIS) to get quantitative and qualitative measures and determine the lowest evaluation score with better approximation. By combining fuzzy logic, used to deal with the inaccuracies and uncertainty problems during the evaluation process, and neural network, used to solve the problem of continuous changes in assessment environments with the delivery of adaptive learning content. By using the ANFIS approach that allows accurate prediction of evaluation scores, the proposed method seems more efficient compared to the recent evaluation methodology.

Currently, Decision support systems in dynamic and complex environment involves the use of visual data mining technology for interactive data analysis and visualization. This paper presents a new architecture for designing such systems. The envisaged architecture based on the Multi-Agent System to improve coordination and communication between the different system modules to generate the appropriate solution for a specific problem. In this work, we have applied the proposed architecture to develop visual intelligent clinical decision support system for the fight against nosocomial infections. The developed prototype was evaluated to show the new architecture applicability.

Concentric circles representation has been developed for visualizing the periodic character of temporal data set. It is particularly useful for visually interpreting periodic time-oriented patterns extracted by data mining techniques. However, this requires a cognitive study to support the transformation into the closest mental representation to reality. In this paper, the proposed information visualization tool is enhanced taking into account key human factors for temporal patterns perception and cognition. This allows facilitating visual analysis of data space to make the right decision by exerting a minimum of cognitive load. We based our work on the taxonomy proposed by Bloom of cognitive domain to improve the concentric circles technique.

In this paper we evaluate interactive visualizations generated by existing medical tool used to visualize patients' fixed and temporal data. It has been developed to better understand a large collection of patients' data in the Intensive Care Unit in order to daily prevent the nosocomial infection occurrence. Existing visualization evaluation studies introduce a variety of classical and novel evaluation methods but we need to integrate more intelligent techniques for the evaluation of visual analytics tasks of medical data. The proposed method consists of interpreting and analyzing initial evaluation results using fuzzy logic technique. Such technique allows intelligent analysis of the evaluation results. Our contribution tends to propose a more novel and interesting method to reach high performance in evaluation procedure.

The presence of large quantities of temporal data requires interactive analysis for decision-making. Interactive decision support system (DSS) based on knowledge discovery in databases (KDD) process proves to be useful. Temporal data visualization techniques are used in the KDD stages to increase the user participation as well as its confidence in the result in order to improve the decision support quality. Our applicative context is the fight against nosocomial infections in the intensive care unit.

The last years witnessed a continued growth of the amount of data. The data analysis and exploration has become more and more difficult. So, it seems important to find means to visually represent this flood of data. Information visualization can help any user to get and understand information efficiently and implicate him/her in the data mining process thanks to our perception possibilities. The visualization domain proposes a large number of information visualization techniques which have been developed over the last decade to support the exploration of large data sets. In this paper, we propose a classification of information visualization techniques. We present also each technique, its advantages and disadvantages.

In this paper we propose an approach aiming to integrate human-computer interaction (HCI) aspects in decision support system (DSS) development. We propose an approach combining two methods: one issued from software engineering field (the rational unified process) and the other one from the HCI field (the U model). We have tested our approach in a DSS set up in the healthcare domain: the supervision of nosocomial infections in an intensive care unit.